Materi Pendukung : T0264P19_2
CLUTTER 16: COMPUTER UNDERSTANDING OF
NATURAL SCENES 1
Version 1
Ulf Grenander, Brown University
Division of Applied Mathematics
Providence, 02912 RI
April, 2002
Computer Vision abounds with techniques for improving images, suppression of noise,
enhancing their appearence, coding and transmission and many other similar tasks. This
has resulted in the emergence of a powerful imaging technology of great usefulness and
applicability. A related but distinct research area is automated image understanding:
computer analysis of images from the real world in terms of signi.cant constituent parts
and the interpretation of the images in terms of categories that are meaningful to humans.
”Is there an object of interest in the picture ?” Standard pattern recognition methods are
clearly insu.cient for this purpose; edge detectors, median .lters, smoothers and the like
can possibly be helpful, but are not powerful enough in order to understand real world
images. Realization of this fact led to the creation of Pattern Theory. Starting from the
belief that natural patterns in all their complexity require complex models for their
analysis, detailed representations of human knowledge to describe and interpret the
images acquired from the worlds encountered. Indeed, we need detailed knowledge
representations, tailor made for the particular image ensemble that we want to
understand. Substantial advances have been made in this direction during the last few
years. Special image algebras have been developed and used for inference for many .
yields
(1) microbiology applied to micrographs of mitochondria: Grenander,Miller(1994)
(2) recognition of moving objects using radar sensors: Lanterman (2000)
(3) analysis of images obtained with FLIR: Lanterman, Miller, Snyder(1996)
(4) hand anatomy from visible light images: Grenander, Chow, Christensen,Rabbit,Miller
(1996) Keenan (1991)
(5) brain anatomy using magnetic resonance images: Christensen,Rabbit,Miller (1996),
Christensen,Rabbit,Miller (1996), Joshi, Miller (2000),Matejic (1998), 1Supported by
NSF-DMS 00774276
1
Wang, Joshi, Miller, Csernansky (2001) (6) pattern analysis in natural scenes: Huang,Lee
(1999), Lee, Mumford, Huang (2001), Lee, Pedersen, Mumford (2001), Pedersen, Lee
(2001) In particular (6) is directly related to the subject of the present paper. In Sections 2
and 3 we shall continue these studies for the particular case of forest scenes. But .rst
consider two models for scenes in general.
1 Pattern Theoretic Representations of Natural
Scenes
We shall consider two models of natural scenes: The Transported Generator
Model, TGM, and the Basic 3D Model, B3M. The .rst one is given by the scene
representation in terms of con.gurations
c = {sνg. . C; ν = ...,-1, 01, ...} (1)
The generators g.(·) are drawn i.i.d from a generator space G and the s’s is a
realization of a stochastic point process over a similarity group, s. . S. This is
simply a mathematical formalizations of the dictum that the material world is
made up of objects and that pattern analysis should be based on this fact. For
pattern theoretic notions see Grenander (1993), referred to as GPT.
An important special case of the TGMis when the similarity group S consists
of translations x. = (x.1, x.2) . R2 in the plane and the con.guration space C is made into
an image algebra, I = C[R], by the identi.cation rule R = ”add”,
see GPT, p. 52, so that
I(x) = 8
_
n = -8
A.g.(x - x. ); I . I; x . R2; x. = (x1ν, x2ν ) . R2 (2)
where the amplitudes A. are iid N(0, s2) and the generators g(·) describe the
shape of the objects in the scene, deterministic or random. The points x. form
a Poisson process in the plane with an intensity λ.
A theorem proved in Grenander(1999a,1990b) says that images governed
by the TGM in (2) has the property that linear operators TI operating on
the resulting image algebra satisfy approximately the Bessel K Hypothesis: 1D
marginal probability densities are of the form
f(x) = 1
vp G(p)cp/2+1/4 xp-1/2Kp-1/2(_2/c|x|) (3)
where K is a Bessel function. It holds for operators T that annihilate constants
and under mild regularity conditions on the g’s.
Another instance of the TGM is when we use the identi.cation rule R =
”min” leading to the image algebra with elements
I(x) = min[A.g.(x - x.)] (4)
2
This version of the TGM corresponds to the situation when scene images are
acquired with a range camera, for example a laser radar.
But the TGM is an extreme simpli.cation of real world scenes. First, it
is 2D rather than 3D. Second, related to the .rst one, it neglects perspective
transformations of objects, and third, it uses a stationary stochastic process
rather than a space heterogenous one. This makes it even more surprising that
the approximation is very precise in most cases as was shown in Srivastava, Liu,
Grenander (2001). Indeed, it is seldom one .nds such close agreement of theory
with data outside of physics and chemistry. This remarkable results makes it
plausible that the TGM can be used with some hope of success also for pattern
analysis of range (laser radar) images. We shall attempt this in Section 2.
There is a need for .rmer support for deriving of algorithms for the recognition
of Objects Of Interest (OOI) hidden against a background of natural scene
clutter, perhaps hiding part of the OOI. This is o.ered by the B3M
scene = .νs.g. (5)
with the objects represented by generator templates g. . Gaν; see GPT p.3,
and, again, the s’s form a stochastic point process over the group S. Here α is
a generator index, see GPT p. 19, that divides the generator space into index
classes
G = .αGa (6)
The generators here mean the surface of the respective objects, and the index
classes could represent di.erent types of objects, trees, buildings, vehicles...
In the case of range pictures it is natural to introduce a 3D polar coordinate
system (r, f, .) where r means the distance from the camera to a point in space
and φ is the azimuth angle and ψ the elevation angle so that we have the usual
relation
x1 = rcosφcosψ; x2 = rsinφconsψ; x3 = rsinψ (7)
A point x = (x1, x2, x3) is transformed into Cartesian coordinates u = (u1, u2)
in the focal plane U of the camera by a perspective transformation that we shall
call T . Hence the range image has the pixel values, in the absence of noise,
I(u) = min.{(Ts.ga. )(u)} (8)
This version of the B3M will be used in Section 3.
1.1 Information Status for Image understanding.
It would be a serious mistake to think of scene understanding as a problem with
the observer equipped with objective and static knowledge about the world from
which the scene is selected. On the contrary, the knowledge, codi.ed into a prior
3
measure, evolves over time and may be di.erent for di.erent observers. The well
known visual illusions speak to this; the ambiguities are resolved completely only
when additional information about the scene is made available.
Think of a person looking for an OOI in a landscape never before see by him
- he will be capable of less powerful inference than some one familiar with the
landscape. If we send out a robot to search for vehicles in a forest it is clear
that it will perform better if equipped with an adequate map than it would
otherwise. This is obvious, but with further reaching implications than may be
thought at .rst glance.
The Hungarian probabilist Alfred Renyi used to emphasize that all probabilities
are conditional. We believe in this, and conclude that any realistic
approach to Bayesian scene understanding must be based on prior probabilities
that mirror the current information status. The automatic search in a desert
landscape for a vehicle using a photograph taken a day ago will be based on a
prior di.erent from the situation with no such photograph, just the knowledge
that it is a desert. In the latter case the prior may be a 2D Gaussian stochastic
process with parameters characteristic for deserts in that part of the world. In
the .rst the prior may be given via a map computed from the photograph superimposed
with another Gaussian processing representing possible changes in
the location of the sand dunes during the last day; obviously a situation more
favorable for the inference machine.
Other incidentals that could/should in.uence the choice of prior are, meteorological
conditions, observed or predicted, position of sun, type of vegetation,
topographical features known or given statistically, presence of artifacts like
camou.age, ... For each likely information status we should build knowledge
informations of the resulting scenes. This is a tall order, a task that will require
mathematical skills and subject matter insight. It should be attempted and it
will be attempted!
1.2 Attention Fields.
A consequence of changes in information status is the e.ect on the prior probability
measure. As more and more detailed information becomes available about
the scene the prior is strengthened. One important instance of this is the atterntion
field that codi.es information about the OOI. It comes in two forms:
(i) For the TGM the attention .eld AF is a probability measure over S
operating on the OOI in the U image plane; it describes the location of the
target as seen by the observer (camera).
(ii) For the B3D representation the attention .eld AF is a probability measure
over the similarity group S operating in the background space X; it describes
location and pose of the target in 3D coordinates.
The AF information can be acquired in di.erent ways. It can come from
4
visual inputs, the operator sees what looks as a suspicious object. It can be the
result of intelligence reports,perhaps with low credibility and accuracy, concerning
the location of the OOI. Or without human intervention, say as the output
of a secondary device like FLIR: the OOI emits heat and the FLIR obtains
crude estimates of its coordinates. Many other technologies can improve the
information status.
2 Pattern Understanding of Forest Scenes.
Let us use the TGM in this section. To create a knowledge representation of
forest scenes we .rst have to decide how detailed it should be. Should it involve
individual trees and bushes? Must branches be speci.ed? Should tree type be
described in the representation, oak, maple, pine...? It all depends upon the
goal we are set for using the representation.
If the goal is to discover man made Objects Of Interest, OOI, say vehicles
or buildings, we may not need a very detailed descripton of the forest. On the
other hand, if the purpose is to automate the collection of tree type statistics
we should include tree type information in the knowledge representation. This
is not just segmentation, it involves analysis and understanding of the content
in the image.
Let us deal with the second of the two alternatives. With some arbitrariness
we have chosen the following four generator indices:
A) ground surface in the foreground, α=”ground”
B) sky background, α = ”sky”
C) trunk element for individual trees, α = ”trunk”
D) close foliage, α = ”foliage”
Now we must make the α-de.nitions precise. Since the TGM is 2D and lives
in the image plane, the generators consist of areas in this plane. We therefore
introduce index operators Oa mapping sub-images IX = {I(x); x . X}, where
X is a subset, say a rectangle, in the image plane,
Oa : IX _. {TRUE,FALSE} (9)
In other words, the index operators are decision functions taking the value
TRUE if we decide that the area X is (mainly) covered by a generator g . Ga.
It may happen that a set X is classi.ed as more than one α-value. We shall order
the way we apply the index operators, one after the other, with the convention
that a TRUE value overwrites previous truth values. We have used the order
5
ground, foliage, trunk, sky.
In tabular form:
Generator Index Classes
Generator Index Informal De.nition Formal De.nition
α = ground Smooth Surface {X:Oground[I]=1}
α = foliage Irregular Surface {X:Ofoliage [I]=1} α = trunk Narrow Vertical Surface {X:
Otrunk[I]=1} α = sky In.nite Distance {X: Osky[I]=1}
2.1 Index operator for”ground”
A ground area is usually fairly .at except for the presence of boulders and low
vegetation like bushes. We shall formalize this by asking that the gradient be
small
Oground[I(X)] = TRUE . _grad(I)(x)_ < c1 = 0; x . X (10)
2.2 Index operator for”foliage”
For foliage on the other hand, the leaves give rise to local variation but this
variation is moderate as long as the leaves belong to the same tree. Hence we
introduce
Ofoliage[I(X)] = TRUE . c2 < _grad(I)(x)_ < c3 = 0; x . X (11)
2.3 Index operator for”trunk”
Trunk areas are narrow and vertical with small local variation compared to their
immediate environment. In Figure 1 the narrow rectangle Xcenter should correspond
to a part of a trunk, it is surrounded by two somewhat larger rectangles
Xleft and Xright with some overlap. Compute the variances of the pixel values
belonging to these three sub-images, call them V arleft, V arcenter, V arright and
de.ne the index operator
Otrunk[I(X)] = TRUE . V arleft > c4V arcenter and
V arright > c4V arcenter; x . X (12)
6
Figure 1
2.4 Index operator for”sky”
This is the easiest generator class to detect. Indeed, sky means in.nite distance,
far away, or rather the largest distance that the laser radar can register. For
the camera used, this is coded as I(x) = 0. Hence we can simply de.ne
Osky[I(X)] = TRUE . I(x) = 0; x . X (13)
7
2.5 Application to Range Images
. Applying this to laser radar images from Lee-Huang(2000):Brown Image Data
Base, we get the following analysis with the meanings displayed graphically.
Figure 2
One sees the occurrence of blue diamonds, ”sky” and two trunks, some
smaller trunk elements too, and a lot of foliage. At the bottom of the .gure is the
ground, separated from the foliage by pixels that could not be understood by the
algorithm. Note that the tombstones in the observed scene have been interpreted
as foliage: the present knowledge representation does not ”understand” tomb
stones. We shall study the understanding of such OOI’s in Section 3. In Figure
8
3 the dominating feature of the analysis is the occurrence of two big trunks.
Figure 3
9
The analysis of Figure 4 also is dominated by a trunk.
Figure 4
10
The analysis of Figure 5 has a sky element again.
Figure 5
It seems that this automated image understanding performs fairly well considering
the primitive form of the index operators that we have used. As we have
argued repeatedly before, a major task in the creation of tools for the automate
understanding of image ensembles is to build specific, tailor made knowledge
representations in pattern theoretic form for the image ensembles in question.
This is a real challenge, one that has been successfully implemented only for a
few scene ensembles, but is a sine qua non for any serious attempt to automate
scene understanding.
11
3 Recognizing Objects Of Interest with Forest
Clutter Background.
When we turn to the problem of automated recognition of OOI’s against a
background of clutter, say a forest scene, the role of the background changes.
In the previous section the background, the forest, was of primary interest, but
now it plays the role of nuisance parameter in statistical parlance. We are now
not after an analysis into foliage, trunks..., but cannot neglect the clutter as
irrelevant. It has long been know that the randomness of clutter is far from
simple white noise. Further, it cannot be modelled as a stationary Gaussian
process in the plane. Indeed the validity of the Bessel K hypothesis contradicts
such a model from the very beginning. We shall use the TGM for representing
the background clutter - the secondary element of the images - and the more
detailed B3M for desribing the OOI’s - the primary element. For the OOI we
choose tanks; we happen to have available a template library in the form of
CAD drawings with rotation angles equal to 0,5,10,15,20... degrees. Since we
position them on the ground we will have s3 = 0, so that this coordinate can
be left out in the computations.
Hence the image representation will take the form
I(u) = min{Iclutter(u), TsgOOI(u); u . image plane U} + e(u) (14)
for the time being with only a single α-value for the OOI, and e(·) standing for
the camera noise. For laser radars the noise level is low, perhaps white noise
e = N(0, s2), s2 << 1. Further the clutter part of the image will be represented
as
Iclutter =_
ν
A.g(u - u.) (15)
We know some approximate probabilistic properties of such representations,
Grenander(1999 a,b).
3.1 Half Bayesian Inference for Range Images
Denote the observed range image by ID(u). If the noise level is low ID(u) ?I(u)
for u-points not obscured by the OOI.With the usual Bayesian reasoning applied
to the present similarity group we shall, however, not assume any prior for the
background; that will be done in the next section. Let us deal with a posterior
density p(s|ID(·) of the form
p(s|ID(·)) . π(s)exp{-1}
2σ2 _ID - min(ID, TsgOOI)_2 (16)
with a prior density π(·) on the similarity group S expressing the current information
status. We admit that it is not clear to what degree this approximation
is adequate for inference.
We shall choose the attention .eld AF as a rectangle in the (s1, s2)-plane
with the center at the pointer (s1AF , s2F ) and with some width (w1AF, w2AF ),
12
preferably biased bised as mentioned below. Let us try a prior of the form
π(s) . exp[(s1 - ponter(1))2
σ2
1(s2 - ponter(2))2
σ2
2(s3 - ponter(3))2
σ2
3
; (s1, s2) . AF; 0 else (17)
not depending on s4, so that the three .rst components are independent and
Gausian when restricted to AF, while the fourth one is uniform on T.
Now search for the MAP estimator. Starting at s = pointer and using the
Nelder-Mead algorithm, see Nelder,Mead (1965), for function minimization, we
will get convergence to a local mimimum, hopefully close to the true s-value, at
least if width is small enough. But the behavior is a bit puzzling. Sometimes it
works well, sometimes not. Why is this so?
One reason is the occurrence of the min operation in (16), typical for range
cameras. If we let s2, the distance away from the camera, be large, the OOI will
eventually be hidden behind the clutter, so that ID - min(ID, TsgOOI) = 0
and only the π factor will play a role; the information in the observed image
will have been wasted, and we will get misleading inferences.
REMARK 1. It follows that if we use straight maximum likelihood estimation
without constraints, the ML estimator is not consistent.
Further, if most or all of the OOI is hidden by the clutter we cannot expect
good inference. On the other hand, if only part of the OOI is hidden, the inference
algorithm should work. This in contrast to methods that are not designed
to take care of obscuration e.ects, for example simple correlation detectors.
REMARK 2. A technical di.culty is that during the computation all the
components of the similarity group element are discrete. This is not serious
however; it can be dealt with.
In order to make the inference algorithm work we should, as mentioned
above, introduce a bias in the choice of the pointer favoring smaller values of
s2. If we do this the inferences are often quite good. We have found it better,
however, to search discretely in the domain of the AP on a fairly coarse grid.
More precisely, we .rst search the translation components and the, keeping the
translation .xed, search over s4. Although we have not looked for fast algoritms
this one seems adequate considering that it is quite crude.
We now apply the algorithm to the range images and display the result by
showing a subset of the image containing the OOI and also the same subset
with the result of the algorithm. In each case the AF was chosen so that it
covered the OOI but with some shift in the ”pointer” away from the center of
the object; this to correspond to mistakes in the information status. We get
13
Figure 6
The inference looks good. Also for Figure 7
14
The same is true for the partly hidden OOI in Figure 8
Figure 8
but the algorithm fails of course in the case of the wholly hidden OOI in
15
Figure 9 and does not discover the OOI
Figure 9
Once the OOI has been found in terms of location and pose the way we
have described, it should be possible to implement a scheme of understanding
by running the algorithm for several di.erent template libraries, to see which
type of OOI we have in the observed image. Lacking an alternative template
library we have not implemented this, and just note that the recognition must
take into account the variability of the window of decision (the subimage used
for recognition) due to variations in perspective when s2 varies.
The performance of this algorithm seems acceptable in these cases, but we
do not know if this is so in general. It is clear that a theoretical study of its
asymptotic properties is needed.
16
3.2 Fully Bayesian Inference for Range Images.
While the algorithm performed fairly well it is likely that it could be improved
by including a prior P over the image algebra I, so that we would adopt a fully
Bayesian point of view, assuming of course that we have access to a reliable
prior I. Then (16) changes into
p(s, I|ID(·) . P(I)π(s)exp{1
2σ2 _ID - min(ID, TsgOOI)_2 (18)
with the prior density P(I); I . I. Using results on the Bessel K hypothesis we
suggest the form
P(I) . _
x.AF
b[I(x1 + 1, x2) - I(x1, x2); phor, chor] ?b[I(x,x2 + 1) - I(x1, x2); pvert, cvert] (19)
with p’s and c’s meaning the parameters in the Bessel K formalism. To make
(19) meaningful we must also specify the boundary conditions on AF. A special
case of some interest is p = 1, the double exponential density
b(y; 1, c) . exp[-y/c] (20)
The resulting, fully Bayesian, recognition algorithm would then compute
the minimum of p(s, I|ID(·) over S ?I say starting with I = ID, and with
s corresponding to the value of ”pointer”. It is easy to write down the Euler
equations for I and s, but we have not yet implemented this procedure in code,
so we do not yet know if it performs beter than the one in Section 3.1.
4 Selected References
G. Christensen, R. Rabbit, M. Miller(1996): Deformable Templates Using Large
Deformation Kinematics. IEEE Transactions on Image Processing, 5(10)
G. E. Christensen, S. C. Joshi, M. I. Miller(1997): Volumetric Transformation
of Brain Anatomy IEEE Transactions on Medical Imaging, vol. 16, pp.
864-877
B. Gidas, A. Lee, D. Mumford: Mathematical Models of Generic Images
U. Grenander, A. Srivastava, M. Miller(2000): Asymptotic Performance
Analysis of Bayesian Object Recognition, IEEE Trans. on Information Theory,
vol. 46, no. 4, pages 1658-1665
U. Grenander, Y.S. Chow, D. Keenan (1991); HANDS, A Pattern Theoretic
Study of Biological Shapes, Springe-Verlag
U. Grenander, M. I. Miller(1994):Representation of Knowledge In Complex
Systems (with discussion section),” Journal of the Royal Statitical Society, vol.
17
56, No. 4, pp. 569-603
U. Grenander(1993): General Pattern Theory, Oxford University Press
U. Grenander(1999a): Clutter 5, Tech.Rep.,www.dam.brown.edu/ptg
U. Grenander(1999b): Clutter 6, Tech.Rep.,www.dam.brown.edu/ptg
J. Huang and A.B. Lee (1999): Range Image Data Base, Metric Pattern
Theory Collaborative, Brown University, Providence RI.
S. Joshi and M. I. Miller(2000): Landmark Matching Via Large Deformation
Di.eomorphisms IEEE Transactions on Image Processing, Vol. 9, No. 8, pages
1357-1370
A. Lanterman(2000): Air to Ground ATT, www.cis.jhu.edu/wu-research/airtoground.html
A Lanterman, M.I. Miller, D.L. Snyder(1996): Representations of thermodynamic
variability in the automated understanding of FLIR scenes,” Automatic
Object Recognition VI, Proc. SPIE VOL. 2756, 26-37
A. Lee, J. Huang(2000): Brown Range Image Database, www.dam.brown.edu/ptg/brid
A. Lee, D. Mumford, J. Huang (2001): ”Occlusion Models for Natural Images:
A Statistical Study of a Scale-Invariant Dead LeavesModel”, International
Journal of Computer Vision, 41(1/2): 35-59.
A. Lee, K. Pedersen, D. Mumford (2001) ”The Complex Statistics of HighContrast Patches in Natural Images”, In Proc. of IEEEWorkshop on Statistical
and Computational Theories of Vision, ICCV 2001, Vancouver, CA,
L.Matejic:( 1998): Testing for brain anomalies: A hippocampus study. Journal
of Applied Statistics, 25(5):593-600
M.I. Miller, S.C. Joshi, D.R. Ma.tt, J.G. McNally, U. Grenander(1994):Membranes,
Mitochondria, and Amoebae: 1, 2, and 3 Dimensional Shape Models, in Advances
in Applied Statistics, (ed. K. Mardia), Oxford, England: Carfax, vol.
II, pp. 141-163
D. Mumford(1966): Pattern theory: a unifying perspective, in ”Proc. 1st
European Congress of Mathematics”, Birkhauser-Boston, 1994, and in revised
form in ”Perception as Bayesian Inference”, edited by D.Knill and W.Richards,
Cambridge Univ. Press.
J.A. Nelder, R. Mead(1965): A simplex method for function minimization,
J. Comp., 7, 308-313
K.S. Pedersen, A.B. Lee (2001): Toward a Full Probability Model of Edges
in Natural Images. APPTS Report 02-1, DAM Brown University, to appear in
the ECCV Proceedings, Copenhagen, May 28-31, 2002
A. Srivastava, X. Liu and U. Grenander. Universal analytical forms for
modeling image probabilities,,www.dam.brown.edu/ptg
L. Wang, S. Joshi, M. Miller, J. Csernansky(2001):, ”Statistical Analysis of
18
Hippocampal Asymmetry in Schizophrenia,” NeuroImage, 14, 531-545
19